Control Chart Recognition Method Based on Transfer Learning

摘要

Control chart is a useful method to identify process problems by detecting abnormal pattern, to keep process in control. Aiming at the problem that traditional abnormal patterns were fixed, and could not completely identify all process problems, a control chart recognition based on transfer learning was proposed. Firstly, six kinds of abnormal patterns were abstracted from the characteristics of abnormal processes. Apply Monte Carlo to create simulated data, and process data to reduce the effect of noise by standardizing and coding. Control chart images were used as target dataset instead of numerical data. According to the feature based transfer learning in the isomorphic space, apply VGG16 network model to extract feature to improve the generalization ability. Finally, the output of the feature extraction was the input of the classifier that has been well trained on the target dataset. Control chart recognition model was fine-tuned according to the recognition results during the process of training, to gain the optimal one. The experimental results show that compared with BP network model, the accuracy of control chart recognition network model based on transfer learning is more than 98% with the less sample data.